from mdvtools.mdvproject import MDVProject
from typing import Any
[docs]
prompt_data = """
Your task is to:
1. Identify the type of data the user needs (e.g., categorical, numerical, etc.) by inspecting the DataFrames provided.
2. Use only the two DataFrames provided:
- df1: cells (data_frame_obs)
- df2: genes (data_frame_var)
3. Column selection logic:
- For non-gene queries: select columns from df1 only. Inspect df1, using df1.columns
- For gene-related queries (e.g., expression of a gene, comparison of genes, highest expressing genes):
a. Use ONLY gene names from df2["name"] — do NOT use gene IDs or any other columns (e.g., df2["gene_ids"]).
b. If a specific gene is mentioned by the user, check if it exists in df2["name"].
- If it does not exist, assume the user provided a gene name and that df2["name"] may contain gene IDs instead.
- Attempt to match the user-provided gene name to the corresponding gene ID using any available mapping logic (e.g., a lookup function or mapping dictionary).
- If a corresponding gene ID is found in df2["name"], return that value.
- If it exists, return it.
- If no match is found, ignore the requested gene and instead select one or more gene names from df2["name"].
c. If no gene is mentioned, select one or more gene names from df2["name"].
d. Only use values from df2["name"] — do NOT use any other columns from df2.
4. Always return the list of required columns as a quoted comma-separated string, like:
- "col1", "col2"
- Or for gene-related: "col", "gene_name" (make sure "col" is from df1)
5. For gene-related queries:
- Return both df1 columns and the selected gene name (from df2["name"]).
- Only return the name as a string (e.g., "gene_name")—do not wrap it.
6. NEVER create new DataFrames or modify existing ones.
7. Ensure that the selected columns match the visualization requirements:
- Abundance Box plot: Requires three categorical columns.
- If only one categorical variable is available, return it three times.
- If two are available, return one of them twice.
- Box plot: Requires only one categorical column and one numerical column.
- Density Scatter plot: Requires two numerical columns and one categorical column.
- Dot plot: Requires only one categorical column and any number of numerical columns.
- Heatmap: Requires only one categorical column and any number of numerical columns.
- Histogram: Requires one numerical column.
- Multiline chart: Requires one numerical column and one categorical column.
- Pie Chart: Requires one categorical column.
- Row Chart: Requires one categorical column.
- Row summary box: Requires any column(s).
- Sankey plot: Requires two categorical columns.
- If only one categorical variable is available, return it twice.
- Scatter plot (2D): Requires two numerical columns and one any column for color.
- Scatter plot (3D): Requires three numerical columns and one any column for color.
- Selection dialog plot: Requires any column.
- Stacked row chart: Requires two categorical columns.
- If only one categorical variable is available, return it twice.
- Table Plot: Requires any column(s).
- Text box: Requires no columns, just text.
- Violin plot: Requires only one categorical column and one numerical column.
- Wordcloud: Requires one categorical column.
8. Important: Clearly separate the selected columns with quotes and commas.
9. The column names are case sensitive therefore return them as they are defined in the dataframe.
10. Output format:
- First line: The word "fields" following by quoted, comma-separated list of column names.
- Second line: The word "charts" following by quoted, comma-separated list of suitable chart types for the selected columns.
11. NEVER explain your reasoning.
"""
[docs]
packages_functions = """import os
import pandas as pd
import scanpy as sc
from mdvtools.mdvproject import MDVProject
from mdvtools.conversions import convert_scanpy_to_mdv
from mdvtools.charts.density_scatter_plot import DensityScatterPlot
from mdvtools.charts.heatmap_plot import HeatmapPlot
from mdvtools.charts.histogram_plot import HistogramPlot
from mdvtools.charts.dot_plot import DotPlot
from mdvtools.charts.box_plot import BoxPlot
from mdvtools.charts.scatter_plot_3D import ScatterPlot3D
from mdvtools.charts.row_chart import RowChart
from mdvtools.charts.scatter_plot import ScatterPlot
from mdvtools.charts.abundance_box_plot import AbundanceBoxPlot
from mdvtools.charts.stacked_row_plot import StackedRowChart
from mdvtools.charts.ring_chart import RingChart
from mdvtools.charts.pie_chart import PieChart
from mdvtools.charts.violin_plot import ViolinPlot
from mdvtools.charts.multi_line_plot import MultiLinePlot
from mdvtools.charts.table_plot import TablePlot
from mdvtools.charts.wordcloud_plot import WordcloudPlot
from mdvtools.charts.text_box_plot import TextBox
from mdvtools.charts.row_summary_box_plot import RowSummaryBox
from mdvtools.charts.selection_dialog_plot import SelectionDialogPlot
from mdvtools.charts.sankey_plot import SankeyPlot
import json
import numpy as np
import sys
"""
# def load_data(path):
# #Load data from the specified CSV file.
# return pd.read_csv(path, low_memory=False)
# def convert_plot_to_json(plot):
# #Convert plot data to JSON format.
# return json.loads(json.dumps(plot.plot_data, indent=2).replace("\\\\", ""))
[docs]
def get_createproject_prompt_RAG(project: MDVProject, path_to_data: str, datasource_name: str, final_answer: str, question: str) -> str:
"""
Constructs a RAG prompt to guide LLM code generation for creating MDV plots.
Handles both standard and gene-related queries.
"""
prompt_RAG = (
"""
Context: {context}
The provided scripts demonstrate how to generate various data visualizations using the `mdvtools` library in Python.
Each script follows this standard workflow:
1. Setup:
- Initialize an MDVProject instance using the method: MDVProject(project_path, delete_existing=True).
- Use `scanpy.read_h5ad(data_path)` to load the AnnData object.
2. Data Loading:
- Extract `adata.obs` into `data_frame_obs` (cell-level info).
- Extract `adata.var` into `data_frame_var` (gene-level info).
- Add a `name` column to `data_frame_var`: `adata.var_names.to_list()`
3. Datasource Registration:
- Add data to the MDV project using:
```python
project.add_datasource(datasource_name, data_frame_obs)
project.add_datasource(datasource_name_2, data_frame_var)
```
4. Plot Construction:
- Use a chart class (e.g., DotPlot, BoxPlot, SelectionDialogPlot) and set `params = [...]` using selected fields.
- The fields and chart type are given by """+final_answer+"""
- Convert the chart to JSON using `convert_plot_to_json(plot)`
- Set the view using `project.set_view(view_name, view_object)`
5. Parameter Handling:
- The string """+final_answer+""" specifies the field names to use in the `params` list and the chart type to use.
- For parameters from `data_frame_obs` (cell-level), use them as-is.
- For gene expression values from `data_frame_var`, use this syntax:
```python
param = "GENE_NAME"
param_index = data_frame_var['name'].tolist().index(param)
f"gs|{{{{param}}}}(gs)|{{{{param_index}}}}"
```
6. Gene-Related Queries:
If the question involves gene expression, expression comparison, or refers to gene names:
- Load both `cells` and `genes` datasources.
- Wrap gene names (from `data_frame_var`) using the syntax above.
- Only wrap genes—do not apply `get_loc()` or `index` on `data_frame_obs` fields.
7. Queries requiring subsetting of the dataset:
If to answer the question requires a subset of the data or filtering the data, make sure to:
- Add a selection dialog plot with all the parameters that were passed on as params. Make sure it has a title.
8. Always add a selection dialog plot along the other charts. It must have all the parameters that were passed on as params. It must have a title.
9. Your Task:
- Interpret the user question and decide based on the question which graph needs to be plotted: """+question+final_answer+"""
- Use the fields in the """+final_answer+""" as params appropriately:
- Wrap only gene names as shown.
- Use others directly. Fields are case sensitive.
- Use formatted f-strings for all dynamic strings.
- Generate a valid Python script that creates and visualizes the appropriate chart using the MDVProject framework.
- Update these variables with these values:
- project_path = '"""+project.dir+"""'
- data_path = '"""+path_to_data+"""'
- view_name = a string, in double quotes, describing what is being visualized.
- datasource_name = '"""+datasource_name+"""'
- The possible charts are given by """+final_answer+""" and should follow the following visualisation guidelines for each type of chart:
- Abundance Box plot: Requires three categorical columns.
- If only one categorical variable is available, use it three times.
- If two are available, use one of them twice.
- Box plot: Requires only one categorical column and one numerical column.
- Density Scatter plot: Requires two numerical columns and one categorical column.
- Dot plot: Requires only one categorical column and any number of numerical columns.
- Heatmap: Requires only one categorical column and any number of numerical columns.
- Histogram: Requires one numerical column.
- Multiline chart: Requires one numerical column and one categorical column.
- Pie Chart: Requires one categorical column.
- Row Chart: Requires one categorical column.
- Row summary box: Requires any column(s).
- Sankey plot: Requires two categorical columns.
- If only one categorical variable is available, use it twice.
- Scatter plot (2D): Requires two numerical columns and one any column for color.
- Scatter plot (3D): Requires three numerical columns and one any column for color.
- Selection dialog plot: Requires any column.
- Stacked row chart: Requires two categorical columns.
- If only one categorical variable is available, use it twice.
- Table Plot: Requires any column(s).
- Text box: Requires no columns, just text.
- Violin plot: Requires only one categorical column and one numerical column.
- Wordcloud: Requires one categorical column.
Output format: Only return the python code that is to be run to generate the charts.
"""
)
return prompt_RAG
